Where We Are Now
Today's AI agents excel at well-defined tasks with clear success criteria, reliable tools, and short time horizons. Give an agent a vague objective, a noisy environment, and a multi-week window, and it struggles. This is not failure of imagination — it is an honest accounting of where the technology stands.
The Next Frontier: Long-Horizon Tasks
The defining challenge for the next generation is long-horizon task completion — plans spanning hours, days, or weeks with thousands of individual actions:
- Persistent state surviving restarts and context resets.
- Reliable world models — accurate tracking of what has been done and what the environment looks like now.
- Robust error recovery — diagnosing failures and replanning without human help.
- Trust calibration — knowing when to act autonomously and when to ask a human.
Computer-Using Agents
Claude Computer Use, OpenAI Operator, and Google Project Mariner demonstrated in 2024 that AI can directly control a web browser or desktop. Instead of needing an API for every tool, agents can use any software a human can use. When agents can use arbitrary GUIs, every piece of software ever written becomes a potential tool.
The surface area of what agents can do is about to explode — not because models got smarter, but because they learned to use the tools humans already built.
The Path to AGI
Most leading researchers now believe AGI will be achieved through increasingly capable agents, not scaling alone. Key milestones: agents conducting original scientific research (2025–2027), agents autonomously bootstrapping new agent systems (2026–2028), and agents with novel problem-solving exceeding human expert performance across domains (2027–2030).
Whether this timeline holds, the direction is clear: we are building software that builds software, and the implications for every field of human endeavour are profound.
